Fix review findings on ban bypass, probe parity, eval gap

Harden PROMPT_EVAL v2 with adversarial questions

Docs and comments match code: env defaults, routing, rich path, stale notes (A12, A2-5..A2-9, A2-14, A3-5, A6-8, A7-10, A8-07, A5-11)

Instant dots, buying-advice routing, br handling, plain-fallback cleanup, live-log model updates

Rich Messages sending, review fixes, no-file note hygiene, persona tone